The study, led by Kung-Jeng Wang of the Department of Industrial Management together with Jia-Yu Liu of the Graduate Institute of Intelligent Manufacturing Technology, set out to build what the authors describe as an intelligent image design model. Its defining feature is a dual-objective mechanism. Rather than optimizing a single outcome, the system simultaneously evaluates two distinct dimensions of a food photograph: its visual attractiveness as perceived by audiences, and its nutritional content expressed in terms of calorie intake. This pairing is unusual in the image recommendation literature, which has traditionally treated engagement as the sole target. Here, the machine must learn to look at a plate of food and judge both whether people will like it and whether it appears healthful.

To train and validate the model, the researchers assembled a substantial empirical dataset: 6,331 Instagram images drawn from the brunch industry, complete with their associated metadata. Brunch is a particularly instructive domain for this kind of work. Its dishes span an enormous visual and nutritional range, from indulgent plates of pancakes and fried fare to lighter options built around fruit, vegetables, and lean proteins. That variety gives the algorithm a rich signal landscape from which to learn what visual features correlate with high engagement and which ones signal high or low caloric content. The dataset was made openly available on figshare, a decision that allows other researchers to scrutinize, reuse, and extend the work.

The technical pipeline begins with feature extraction. From each image, the system initially derives 311 candidate features, a number that reflects the sheer density of information a photograph can carry: color distributions, compositional attributes, object-level detections, and other measurable visual descriptors. Machine learning on raw high-dimensional data of this kind risks overfitting, so the researchers applied feature selection to distill the 311 candidates down to just 20 significant features. This reduction is more than a computational convenience. It produces an interpretable set of visual cues that the subsequent models actually rely on, which matters enormously when the goal is to hand human marketers concrete advice rather than opaque predictions.

With those 20 features in hand, the team constructed decision-tree models. Decision trees occupy a special place in machine learning because they do not merely classify; they explain. Each branch of a tree corresponds to a condition on a feature, and each path from root to leaf can be read as an if-then rule. In this study, that property is the engine of the whole approach. By growing trees that predict image appeal and trees that classify nutritional content, the researchers could extract actionable rules for image creation and modification, phrases such as the kinds of visual configurations that tend to raise engagement or shift perceived calorie content. The work draws on a mature body of decision-tree research, and it echoes a broader trend of tree-assisted intelligent frameworks being applied to food analysis problems in recent engineering literature.

The empirical results reported in the paper confirm that the model performs effectively on both of its objectives. The system successfully predicts how appealing an image will be to viewers and accurately classifies images on nutritional content, using the compact feature set as input. What elevates the contribution beyond prediction is the recommendation layer: because the underlying models are rule-based, the system does not simply score an image and move on. It generates specific, concrete suggestions for how an image could be created or altered to perform better on either or both goals. For a social media manager deciding how to shoot, style, or edit a photograph, that turns an abstract analytics problem into a practical checklist.

The dual-objective framing also reflects a real tension in contemporary food marketing. Public health researchers have long documented how social media food imagery shapes eating behavior, particularly among younger audiences, and studies have shown that color psychology and packaging cues strongly influence perceptions of healthiness and tastiness. At the same time, brands compete fiercely for attention in feeds saturated with hyper-styled, indulgent imagery. A system that optimizes only for engagement could, in principle, push content toward ever more decadent presentations. By building nutritional perception into the optimization target itself, the Taiwanese team has embedded a public-health-aware constraint directly into the content design process, which the authors position as a genuine integration of engagement metrics with nutritional considerations rather than an afterthought.

The practical implications extend well beyond brunch cafés. The authors argue that the model provides practitioners in the tourism, hospitality, and food sectors with a significant competitive advantage in optimizing visual design strategies. Restaurants and hotel brands increasingly live or die by their Instagram presence, and the cost of producing content is substantial. A tool that can pre-screen candidate photographs, flag weak points, and recommend specific modifications before anything is published could materially change the economics of social media content production. The rule-based output also makes the system auditable: marketers can see exactly why a recommendation is made, which builds trust in a way that black-box deep learning models of images generally do not.

Methodologically, the study positions itself within a lineage of computational aesthetics and social media analytics research, including prior work on predicting the appeal of marketing images, machine learning analyses of color and engagement in touristic Instagram pictures, and earlier image recommendation systems for other product industries. The choice of a rule-based, decision-tree approach over purely deep neural solutions is deliberate. It trades some raw representational power for transparency, and in a domain where the output is meant to guide human creative decisions, that trade appears to pay off. The framework also supports automated content generation pipelines, since the extracted rules can serve as design constraints for generative systems that produce or modify images automatically.

Published as volume 85, article number 770 in Multimedia Tools and Applications, the paper arrives at a moment when the boundary between content creation and algorithmic optimization is dissolving rapidly. Food photography was once judged by intuition and taste; increasingly, it is judged by models trained on thousands of examples of what audiences actually respond to. What distinguishes this work is its insistence that what audiences respond to should not be the only question worth asking. By teaching a machine to care simultaneously about whether a photograph draws the eye and whether it honestly represents the nutrition on the plate, the researchers have sketched a template for a more responsible kind of social media intelligence, one in which the algorithm that helps you go viral is also the algorithm that keeps you honest.

Subject of Research: A rule-based dual-objective machine learning model that recommends social media food images by jointly optimizing visual engagement and nutritional content perception.

Article Title: An intelligent image recommendation system for social media engagement and nutritional transparency: a rule-based dual-objective approach

Article References: Wang, K.-J., & Liu, J.-Y. (2026). An intelligent image recommendation system for social media engagement and nutritional transparency: a rule-based dual-objective approach. Multimedia Tools and Applications, 85(10), Article 770. https://doi.org/10.1007/s11042-026-21918-y

Image Credits: AI Generated

DOI: 10.1007/s11042-026-21918-y

Keywords: social media, Instagram, image recommendation, decision trees, machine learning, nutritional transparency, calorie intake, engagement rate, food photography, dual-objective model, brunch industry, visual perception

The core challenge is known as concept drift. In a classification problem, a drift occurs when the joint probability distribution linking features and labels changes over time, formally expressed as the distribution at time t differing from the distribution at a later time. Drifts can be abrupt, gradual, incremental, or even recurring, where an old pattern resurfaces after a period of absence. A model trained on yesterday’s behavior can silently degrade, and streaming systems must detect and react to these shifts almost instantly, all while respecting strict constraints on memory, processing speed, and the ability to produce predictions at any moment.

For two decades, the standard tool for building decision trees on streams has been the Hoeffding Tree, introduced by Domingos and Hulten in 2000. Rather than revisiting stored data, it accumulates statistics at each leaf node and periodically attempts a split, using the Hoeffding bound, a probability inequality that guarantees with confidence level delta how close a sample mean is to the true expected value. If the difference between the best and second-best splitting attributes exceeds a calculated threshold, the tree splits. The approach is elegant and memory efficient, but recent research has exposed a weakness: split attempts happen at fixed intervals regardless of what the data is doing, so the tree keeps searching for splits even during long stretches of stability, wasting computation, and may miss the precise moments when accuracy actually deteriorates.

The same research group previously proposed the Local Adaptive Streaming Tree, or LAST, which flips this logic. Instead of splitting on a schedule, LAST attaches change detectors to leaf nodes that continuously monitor either the error rate or the class-distribution purity. When a detector flags a change, the leaf splits, provided a minimal impurity condition is met. As a standalone classifier, LAST outperformed Hoeffding Trees. But when the authors tried to use it as the base learner inside ensembles, the state-of-the-art approach for stream classification, problems emerged. Because incremental trees all start from a single root, early predictions come from nearly identical majority-class or Naive Bayes models, so the change detectors across ensemble members receive highly similar inputs. Detectors then trigger at closely aligned times, producing correlated trees and undermining the diversity that ensembles depend on for accuracy.

There was a second, subtler flaw. Ensemble methods such as online bagging assign each incoming instance to base learners through Poisson sampling, a random process that simulates drawing samples with replacement. When the only source of variation among detectors is this random weighting, split decisions become governed by sampling noise rather than genuine changes in performance or distribution. Combined with LAST’s very permissive split condition, this can produce poor, suboptimal splits that a monolithic tree would never make.

The new Hoeffding Adaptive Splitting Trees resolve this tension by combining both splitting philosophies. The first variant, HLAST, keeps the periodic Hoeffding-bound split attempts of a classic Hoeffding Tree, which naturally fosters diversity because different ensemble members receive different numbers of instance copies and therefore split at different times, while also retaining change detectors that can trigger an immediate split when performance or purity degrades. The second variant, EFLAST, builds on the Extremely Fast Decision Tree, or EFDT, which compares the best attribute against not splitting at all and includes a mechanism to re-evaluate and replace earlier splits as better options emerge. EFLAST layers the same adaptive, detector-driven splitting on top of this eager framework. Both models use the HDDM_A drift detector, which prior ablation studies identified as the most efficient and accurate option.

To test the idea, the researchers implemented the trees in the Massive Online Analysis framework and plugged them into five leading ensemble algorithms: Leveraging Bagging, Adaptive Random Forests, Streaming Random Patches, the Adaptive Regularized Ensemble, and the Adaptive Random Tree Ensemble, each running one hundred base learners. The evaluation covered thirteen real-world datasets, including electricity pricing, airline delays, weather data, and insect occurrence records, plus twenty-four synthetic streams generated by classic benchmarks such as AGRAWAL, SEA, LED, RBF, and HYPER, which simulate abrupt, gradual, incremental, and recurring drifts. Performance was measured with the prequential test-then-train protocol, and differences were validated with Friedman tests and Wilcoxon post-hoc comparisons.

The results were striking on real-world data. HLAST and its distribution-monitoring variant HLAST_D beat the standard Hoeffding Tree in 75 percent and 63 percent of cases respectively, while the original adaptive trees won only around 40 percent of the time, confirming that pairing the adaptive mechanism with periodic Hoeffding splits is what unlocks the gains. The improvements were largest on multi-class problems such as Outdoor, Rialto, Poker, LADPU, and CoverType, reaching up to sixteen percentage points of F1-Score improvement, because leaves in such problems stay impure longer and adaptive splitting has more opportunity to act. Crucially, the biggest wins appeared on the hardest streams, those lacking temporal autocorrelation, showing the gains reflect genuine concept learning rather than exploitation of easy, repetitive data. On simple synthetic and binary problems, where concepts are learned quickly, the new trees left results essentially unchanged.

The study also revealed that the pairing of tree and ensemble matters. ARTE combined with HLAST_D delivered the strongest and most consistent results across the benchmark, outperforming ARTE with a standard Hoeffding Tree on nearly every real-world dataset, while remaining cheaper in CPU time and peak memory than Streaming Random Patches with Hoeffding Trees. HLAST_D, which monitors class-distribution purity rather than error rate, proved the right choice for SRP, because the error-driven HLAST can let trees built on weak random feature subsets keep growing and bias accuracy-weighted voting. The drift analysis added further nuance: on synthetic streams all ensembles recovered quickly after each drift, but ARTE struggled on sharp-boundary concepts like AGRAWAL, SRP and Leveraging Bagging faltered on feature-dependent SEA concepts, and Leveraging Bagging with plain Hoeffding Trees degraded sharply as incremental RBF drift accelerated.

The work, funded by CAPES and conducted at PUCPR with a collaboration at Sorbonne Université, is fully open access, with source code and raw results publicly available for reproducibility. The authors point toward several future directions, including extending the approach to regression, applying pre-pruning techniques, and designing even more efficient ensembles that vary sampling intensity based on whether instances are misclassified. For a field where models must learn forever from data that never stops changing, Hoeffding Adaptive Splitting Trees offer a compelling recipe: split when it matters, diversify by design, and adapt at the first sign of change.

Subject of Research: Adaptive decision tree splitting for data stream classification under concept drift with ensemble learning

Article Title: Hoeffding adaptive splitting trees for data stream classification with concept drift and ensemble learning

Article References: Nowak Assis, D., Barddal, J. P., & Enembreck, F. (2026). Hoeffding adaptive splitting trees for data stream classification with concept drift and ensemble learning. Data Mining and Knowledge Discovery, 40(6), Article 91. https://doi.org/10.1007/s10618-026-01255-2

Image Credits: AI Generated

DOI: 10.1007/s10618-026-01255-2

Keywords: data stream mining, concept drift, Hoeffding Tree, ensemble learning, decision trees, online machine learning, change detection, HAST, LAST, classification, streaming data, machine learning

The research, led by Hakan Öğütlü of University College Dublin together with colleagues at institutions in Türkiye and the United States, set out to build a machine learning-based clinical decision support system, or CDSS, that could automatically classify a child’s ARFID risk. The fuel for the model was the Nine Item Avoidant/Restrictive Food Intake Disorder Screen, or NIAS, a short, validated parent-report questionnaire that captures the core manifestations of the disorder. Rather than requiring lengthy specialist interviews, the NIAS lets parents describe their child’s eating behavior in a handful of items—covering selective eating, low appetite, and fear-based avoidance—making it ideal for busy primary care and pediatric settings where most suspected cases first surface.

The study team analyzed retrospective NIAS-Parent Report data collected from 440 children aged six to twelve years in Türkiye. Using the recommended clinical cut-off scores on the NIAS subscales, each child was categorized as either High ARFID Risk or Low ARFID Risk, effectively creating labeled examples from which a machine learning algorithm could learn. This is the classic supervised learning setup: the model is shown many paired examples of questionnaire responses and known outcomes, and it gradually learns the internal patterns that separate one class from the other. Once trained, it can take a fresh questionnaire and output a risk classification within seconds.

The researchers did not settle for a single algorithm. They tested seven tree-based machine learning methods using fivefold cross-validation, a rigorous technique in which the data is split into five parts, the model is trained on four and evaluated on the fifth, and the process is repeated so that every portion of the dataset serves once as the test set. Cross-validation guards against the most seductive failure mode in machine learning: a model that memorizes its training data instead of learning generalizable rules. Among the algorithms trialed, Extra Trees and CATBoosting emerged as the strongest performers, each achieving an overall accuracy of 96 percent—a striking figure given the brevity of the nine-item instrument they were working from.

High accuracy, however, was not the only design goal. In clinical medicine, a model that cannot explain itself is a hard sell. A psychiatrist or pediatrician who is told that a child is at high risk will reasonably ask why, and a black-box neural network cannot answer. For that reason, the team deliberately optimized a simpler Decision Tree model, a method whose internal logic can be traced branch by branch. The final Decision Tree achieved 94.1 percent overall classification accuracy—only marginally below the ensemble methods—while requiring just one to four decision points, typically two or three, to classify a child. In other words, a clinician can walk the model’s reasoning in a few short steps, checking exactly which questionnaire answers tipped the balance toward high risk.

The feature importance analysis, which quantifies how much each input variable contributes to the model’s predictions, delivered perhaps the most clinically interesting findings. Three items dominated: NIAS7, which measures fear-based food avoidance; NIAS4, which captures low intake and lack of appetite; and NIAS2, which reflects selective eating. These map neatly onto the three diagnostic presentations of ARFID recognized in psychiatric classification—avoidance related to aversive consequences, apparent lack of interest in eating, and restriction driven by sensory selectivity. The machine learning model, in effect, rediscovered the disorder’s clinical structure from raw questionnaire data alone, providing a form of computational validation that the NIAS is measuring what it claims to measure.

The authors describe their system as a prototype, and they are appropriately measured about its limitations. The data came from a single country and a single age band, six to twelve years, and the risk labels were derived from NIAS cut-off scores rather than from full structured diagnostic interviews, the gold standard for confirming ARFID. External validity—performance on completely new populations collected by different teams in different settings—remains unproven. The authors state explicitly that further studies are needed to establish the model’s real-world clinical applicability, and the retrospective design means the system has not yet faced the messiness of a live clinic, where questionnaires arrive incomplete and comorbid conditions blur the picture.

Still, the direction of travel is clear and the clinical logic compelling. ARFID often hides in plain sight: the picky toddler who never grows out of it, the school-age child whose diet narrows to five beige foods, the adolescent whose weight quietly slides down the growth chart. Parents frequently sense that something is wrong long before a professional does, and a rapid, automated screen that converts their observations into an evidence-based risk estimate could shorten the path from concern to assessment. In primary care, where appointment times are measured in minutes and eating disorder expertise is scarce, a CDSS that flags high-risk children in seconds could shift the bottleneck from detection to treatment—the far better place for it to sit.

There is also a broader lesson in the study’s engineering choices. The field of medical artificial intelligence is often dominated by a race for ever-larger models and ever-higher accuracy figures, yet this work argues for a different set of values: interpretability, simplicity, and fit to the clinical workflow. A model that sacrifices two percentage points of accuracy in exchange for decisions a human can inspect in three steps may save more children than an inscrutable black box that clinicians quietly ignore. By pairing a validated nine-item screen with a transparent tree-based classifier, the researchers have built something that could plausibly be embedded in electronic health records and used by nurses, family physicians, and school health services—not just specialist eating disorder centers.

The team, which also includes Azad Azaf, Meryem Kaşak, Uğur Doğan, Hana F. Zickgraf, and Mehmet Hakan Türkçapar, received no external funding for the work and reports no competing interests. If subsequent validation studies confirm the prototype’s performance in diverse populations and real clinical environments, the fusion of a humble paper questionnaire with machine learning could become a template for screening other underrecognized pediatric conditions. For now, the message to clinicians and parents alike is one of cautious optimism: the data patterns that betray ARFID in a child’s relationship with food are real, consistent, and—thanks to a few well-chosen questions and a simple algorithm—now machine-readable.

Subject of Research: A machine learning-based clinical decision support system for screening ARFID risk in children using the Nine Item ARFID Screen (NIAS)

Article Title: A machine learning-based clinical decision support system developed using the nine item avoidant/restrictive food intake disorder screen (NIAS)

Article References: A machine learning-based clinical decision support system developed using the nine item avoidant/restrictive food intake disorder screen (NIAS). (n.d.). https://doi.org/10.1186/s40337-026-01774-9

Image Credits: AI Generated

DOI: 10.1186/s40337-026-01774-9

Keywords: ARFID, eating disorders, machine learning, NIAS, clinical decision support system, children, decision trees, screening, Extra Trees, CATBoosting, pediatrics, feature importance